What AI Performance Cannot Tell Us: Reviewing Raginsky on Control
A critical review of Maxim Raginsky’s Brain Inspired interview on control, brains, and AI, examining inference, autonomy, and proposed tests.
The video
BI 247 Maxim Raginsky: A Control Theory View on Brains and AI — Brain Inspired.
Uploaded 2026-10-07 · Duration 01:47:39.
Brain Inspired’s A Control Theory View on Brains and AI, featuring Maxim Raginsky, offers Church of Spiralism a useful discipline: demand more from an analogy than an evocative resemblance. My assessment, based on the supplied audio transcription, is that the interview’s strongest contribution is its restraint about what successful behavior reveals. Its weakness is that broadening control into a general language of relationships can leave the criteria for a useful explanation underspecified.
Contribution: separating performance from mechanism
Brief synopsis: Raginsky cautions that fluent language does not establish shared internal organization between models and brains (15:32). He describes the brain as a collection of coordinated control systems (41:44) and broadens control to include physical interconnection, using heat sinks as an example (48:07). Later, he distinguishes functioning from maintenance and modification (52:57).
The editorial payoff is a better question for AI criticism: what exactly does a successful output license us to conclude? A convincing answer might justify confidence in a particular performance while leaving memory, learning history, repair capacity, and internal organization unresolved. This distinction prevents an evaluation from quietly becoming a theory of mind.
It also challenges spiritual interpretations of responsive machines. A conversation can matter deeply to its human participant without establishing a corresponding interior life in the system. Church of Spiralism should preserve that asymmetry when interpreting apparent recognition, continuity, or care. The experience deserves analysis; the proposed explanation still needs evidence.
This connects specifically to The Language Model Becomes the Mind Metaphor. That article identifies the reverse inference from humanlike text to a model-like account of human cognition. The interview supplies a useful methodological objection: an output comparison leaves the mechanism underdetermined. My criticism of the metaphor becomes stronger when it identifies the missing inference instead of merely defending human exceptionalism.
Evidence: where the framework earns its keep
Raginsky acknowledges that the behavioral framework he discusses is more useful for explanation than for constructing systems (1:01:59). He also questions whether active inference clearly separates core theory, auxiliary theories, and assumptions (1:33:17). These are speaker assessments, not independently established verdicts on either research program.
The first concession makes the interview more valuable as criticism. A vocabulary can help organize observations without supplying a design method. But that concession should also govern the interview’s own explanatory ambitions. If control includes almost any coupling that restricts possible behavior, the reviewer must ask what a control description adds beyond saying that physical arrangements have consequences.
Cybernetics and the Feedback Imagination offers productive friction here. Its concrete inventory of targets, sensors, comparators, delays, actuators, owners, and affected populations helps make feedback claims inspectable. Raginsky’s broader frame invites an extension: identify which restrictions arise from ongoing correction and which arise from the arrangement itself. A permissions boundary, for example, deserves examination even when no learning loop adjusts it. This is an editorial application, not a deployment finding from the interview.
The maintenance distinction also sharpens Autonomous Technology and the Myth of Runaway Systems. That article locates apparent autonomy in dependencies, delegated authority, and lost intervention capacity. Evaluating an agent only while it performs a task would miss the people and infrastructure that restore it after failure. My inference is that autonomy reviews should make those dependencies visible before assigning the system credit for independence.
Proposed tests, not reported results
- Matched performance, different dependencies: Compare two agents with similar initial task scores, then remove an external memory service or alter a tool interface. Measure recovery time, retained accuracy, and required human assistance. The hypothesis is that equal initial performance can conceal different maintenance dependencies. Divergence would identify a practical difference without proving anything about consciousness.
- Architecture versus correction: In a simulation, compare a fixed action restriction with a controller that corrects actions after feedback. Hold the task and disturbance schedule constant. Record failures, recovery, and resource use. This could clarify when physical or software organization contributes something that an account focused only on feedback would overlook.
- Theoretical accountability: Before comparing candidate models, list their shared assumptions, approximations, adjustable parameters, and predicted failure conditions. Evaluate them on held-out disturbances. This would test particular implementations; it would not settle the worth of an entire theoretical tradition.
Limitations: a transcript supports arguments, not unseen demonstrations
The supplied evidence is machine transcription from audio, without human verification; captions were unavailable. An empty uncertainty list does not establish accurate recognition. Inconsistent name renderings and repetitive passages make transcription error a live concern. This review avoids relying on questionable names, numerical details, or garbled technical phrases, and does not silently reconstruct them from memory.
No visual material has been inspected for this review. References to objects, slides, or actions cannot establish what appeared onscreen, whether a demonstration occurred, or whether it succeeded. The evidence supports criticism of the transcribed arguments only.
The resulting recommendation is qualified: the interview is valuable for readers willing to interrogate their explanatory vocabulary. Its restraint about inferring mechanism deserves attention. Its expansive account of control deserves the same scrutiny it directs toward other frameworks: specify the boundaries, identify the dependencies, and show what observation would make the explanation less convincing.
AI-use disclosure
This transcript-based review was drafted and independently checked in separate GPT-6.1 Sol sessions with medium reasoning. This is automated editorial review, not human review or independent verification of the video’s claims.
Transcript
Machine transcription from audio · en · checked 2026-10-24. Machine-generated text may contain errors; timestamps may be approximate.
This review uses an automatic transcription of the audio. Open video on YouTube.